Enhancing the Performance of Sign Language Recognition Models Using Machine Learning

Sign language recognition and related applications heavily rely on image processing. Interpreting and comprehending sign language gestures used by the deaf and hard of hearing is known as sign language recognition. The visual data derived from sign language gestures is analyzed and processed using image processing techniques. Many different machine learning (ML) based approaches have been proposed to accurately detect sign language gestures. In this paper, two distinct pre-processing approaches are introduced for the American Sign Language (ASL) alphabet dataset, namely a threshold-enabled approach and a threshold-less approach. The goal is to investigate the impact of the proposed pre-processing approaches on the performance of the recognition models. The ASL alphabet dataset is utilized to train several ML classifiers including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naive Bayes (NB). Additionally, a combination of SVM and KNN is implemented using ensemble techniques such as bagging and stacking to further improve the performance. An extensive evaluation of the two proposed pre-processing approaches is conducted, considering their impact on multiple performance metrics (accuracy, recall, precision, Fl-score). Experimental results show that the ensemble bagging model that combines SVM and KNN achieves superior performance when using the threshold-enabled preprocessing approach while the SVM model outperforms other classifiers when using the threshold-less pre-processing approach.